EXPLORING THE RELATIONSHIP BETWEEN TRAVEL PATTERN …€¦ · EXPLORING THE RELATIONSHIP BETWEEN TRAVEL PATTERN AND SOCIAL-DEMOGRAPHICS USING SMART CARD DATA AND HOUSEHOLD SURVEY

EXPLORING THE RELATIONSHIP BETWEEN TRAVEL PATTERN AND SOCIAL-

DEMOGRAPHICS USING SMART CARD DATA AND HOUSEHOLD SURVEY

Yang Zhang 1, *, Tao Cheng 1, Nilufer Sari Aslam 1

1 SpaceTimeLab for Big Data Analytics, Dept. of Civil, Environmental & Geomatic Engineering, University College London, Gower

Street, London, WC1E 6BT – (yang.zhang.16, tao.cheng, n.aslam.11)@ucl.ac.uk

Commission IV, WG IV/10

KEY WORDS: Smart card data, Travel pattern, Social-demographics, Household survey, Relationships, Clustering

ABSTRACT:

Understanding social-demographics of passengers in public transit systems is significant for transportation operators and city

planners in many real applications, such as forecasting travel demand and providing personalised transportation service. This paper

develops an entire framework to analyse the relationship between passengers’ movement patterns and social-demographics by using

smart card (SC) data with a household survey. The study first extracts various novel travel features of passengers from SC data,

including spatial, temporal, travel mode and travel frequency features, to identify long-term travel patterns and their seasonality, for

the in-depth understanding of ‘how’ people travel in cities. Leveraging household survey data, we then classify passengers into

several groups based on their social-demographic characteristics, such as age, and working status, to identify the homogeneity of

travellers for understanding ‘who’ travels using public transit. Finally, we explore the significant relationships between the travel

patterns and demographic clusters. This research reveals explicit semantic explanations of ‘why’ passengers exhibit these travel

patterns.

* Corresponding author

1. INTRODUCTION

The portable and durable smart card (SC) has been widely used

for paying for public transport, such as London’s Oyster card

(Lathia et al. 2011), Beijing’s BMAC card (Yuan et al. 2013),

Singapore’s SC for MRT service (Sun et al. 2012). SC that

stores massive trip transactions of passengers has been drawn a

lot of attention in various existing literature (Pelletier et al.

2011). The application domains include mobility pattern

analysis (Shi et al. 2014), traffic congestion pattern analysis

(Ceapa et al. 2012), home/work location estimation (Sari Aslam

et al. 2018), and activity detection (Nassir et al. 2015).

Overwhelming amounts of SC data also provides a promising

way to mine mobility patterns for better transport planning and

service provision. However, it lacks the social-demographic

information of passengers to further explore ‘who are the card

carriers’, ‘why they behaved differently’ and ‘what factors

affect their behaviours’, which are crucial to better understand

the users’ travel demand and mobility patterns. Fortunately,

leveraging household survey data, it might further explore the

relationship between human travel patterns and their social-

demographic roles (Zhang et al. 2018; Zhang et al. 2019),

which can help operators make better transportation planning

and provide passengers with more personalised services.

In this paper, an entire framework is proposed to explore ‘how’,

‘who’ and ‘why’ travels in the PT:

‘How’: We aim to establish an elaborate travel feature

extraction process to classify passengers’ long-term travel

behaviours by using smart card data. Users are then clustered

into several groups indicating different travel patterns for the in-

depth understanding of ‘when’, ‘where’ and ‘how often’

people travel by ‘which travel mode’ in cities.

‘Who’: Leveraging travel survey data, passengers can be also

categorised into different demographic groups based on

individual or household demographic variables, including age,

working status, main occupation, car ownership, household

income. This analysis investigates who usually travel via public

transit (e.g., bus or underground).

‘Why’: In this step, we link the passengers’ travel pattern with

the demographic group to find the significant linkages between

the two clustering results. This study provides a better

understanding and semantic explanations of passengers’

movement patterns.

This paper is organised as follows. Section 2 introduces the

dataset used in this study. Section 3 illustrates the

methodologies to analyse the travel patterns, social-

demographics groups and their relationships. Then, Section 4

describes a case study of London, UK. Finally, the conclusions,

limitations and future work are discussed in Section 5.

2. DATASET

2.1 London’s Oyster Card Data

The SC data used in this study is a sample of Oyster Card

transaction records in London, UK, during the full year of 2012.

There are two types of SCD, one from the tube system and the

other from the bus system. A transaction is recorded

automatically when a passenger taps in/out at a tube station or

boards at a bus stop. Summarily, the entire dataset contains

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-2/W13, 2019 ISPRS Geospatial Week 2019, 10–14 June 2019, Enschede, The Netherlands

This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-2-W13-1375-2019 | © Authors 2019. CC BY 4.0 License.

1375

around 2.18 million journeys made by 9708 passengers, made

up of 33.7% tube journeys and 66.3% bus journeys. Each SCD

record consists of the following fields: (1) Oyster card ID

(encrypted), (2) transaction date, (3) start time, (4) end time, (5)

boarding station, (6) exiting station, (7) journey mode (bus or

tube). Note that in bus trip records, the boarding station

indicates the bus line number but not precise locations, and the

exit station and end time are unavailable.

2.2 London Travel Demand Survey Data

London Travel Demand Survey (LTDS) is a continuous survey

based on the household for collecting individual or household

demographic, social-economic and travel-related information.

Every year, around 8000 randomly selected households

undertake the LTDS annually. All household members aged 5

and over are required to complete the questionnaire. The

information provided in LTDS includes: (1) Oyster card ID (2)

PAGEI: Age, (3) PMANAGER: If a manager, (4) HCVN:

Number of vehicles in total owned, (5) HINCOMEI: household

income, (6) PWKSTAT: working status, (7) POCCUPA:

occupation type, (8) POFWK: weekly work frequency, (9)

PLENN: approximate daily commuting distance, (10)

PFRCARD: the frequency of using car as a driver, (11)

PFRCARP: the frequency of using car as a passenger. Among

them, ‘PAGEI’ and ‘PLENN’ are continuous variables, and

others are categorical variables.

3. METHODOLOGY

3.1 Framework

This article aims to explore ‘how’ (including ‘when’ and

‘where’), ‘who’ and ‘why’ travel in public transit using smart

card data and household survey. For such purpose, the proposed

framework should be capable of:

• Step 1: Identify long-term travel patterns by using smart

card data, telling how passengers travel in the city.

• Step 2: Identify social-demographic groups of travellers,

understanding who travels.

• Step 3: Define more significant relationships between

travel patterns and social-demographic groups.

The framework is illustrated in Figure 1.

Figure 1. Methodology framework

3.2 Travel Pattern Analysis

Traditional travel pattern analysis using smart card data focuses

on daily frequent trip pattern recognition, including (Kieu et al.

2014; Tao et al. 2014), which cannot reflect the full and

trustworthy portraits of passengers during a long-term range,

such as yearly travel pattern. To overcome this issue, the paper

proposes to first distinguish travel patterns using travel features

extracted from SC data. In addition, two novel statistic

measures are employed to identify and quantify the seasonality

of different travel patterns.

3.2.1 Travel Feature Extraction

A key issue in passenger segmentation based on their travel

behaviours is to extract accurate and comprehensive travel

features from SC data. In this study, various travel features are

defined as to calibrate passenger profiles in order to

differentiate their travel patterns. All features are categorised

into four types, related to temporal variability (When), spatial

variability (Where), travel mode preference (Which mode) and

travel frequency (How often), respectively. Authors have

demonstrated and explained the feature extraction process in

(Zhang et al. 2017). Here, we just list the features generated

from SC data in Table 1. The morning and evening peak for

London Underground is between 6:30 and 9:30 and between

16:00 and 19:00 on weekdays, respectively.

3.2.2 Affinity Propagation for Travel Pattern Clustering

In this paper, we propose to use Affinity Propagation (AP)

algorithm for travel pattern clustering. AP, first developed by

Frey et al. (2007), is a local-message-passing-based clustering

approach. It has many advantages in terms of clustering task.

Unlike other clustering algorithms, such as centroid-based k-

means or k-medoids, AP does not require the predefined

number of clusters before running this algorithm. Furthermore,

AP takes all data points as candidates of exemplars (the centre

of cluster). Since we hardly have any prior knowledge about

underlying travel patterns, travel pattern identification can

benefit from the above-mentioned advantages. The details of

AP can be referred to (Frey et al. 2007).

3.2.3 Identify and Quantify Seasonality of Travel Patterns

Seasonal traffic demand may obviously increase the burden on

urban public transit systems. Understanding long-term travel

behaviours will help transportation agencies formulate better

strategies and make more effective and efficient operating

policies. In this paper, we propose two novel statistic measures,

skewness, and kurtosis of trip distributions, to identity and

quantify the seasonality of travel patterns, revealing more

details of passengers’ travel habits.

(1) Seasonality identification

This paper proposes to use the skewness of the trip distribution

by month as a quantitative measure to detect whether a travel

pattern exhibit seasonality. In statistics, skewness is a measure

of the asymmetry in a distribution. Suppose the number of trips

in each month during a year is 1 2, , , Nx x x , the skewness is:

( )

3

1

3

N

iix x N

sks

=−

=

(1)

where x is the mean value, s is the standard deviation, and N is

the sample size. The value of skewness can be positive or

negative. Positive skewness indicates data that are right skewed,

and vice versa. To interpret the values for skewness, Bulmer

(1979) suggests the following rule of thumb:

• If |sk| > 1, the distribution is highly skewed.

• If 0.5< |sk| <1, the distribution is moderately skewed.

• If |sk| < 0.5, the distribution is approximately symmetric.

Hence, if the skewness of a travel pattern’s trip distribution

within (-0.5, 0.5), it is regarded as an unseasonal travel pattern.

Otherwise, the travel pattern should be a seasonal one.



1376

(2) Seasonality quantitation

To quantitative analysis each travel pattern’s preferred seasons

or months for travelling via public transit, we employ a statistic

measure ‘excess kurtosis’ to evaluate the heaviness of the tails

of a distribution relative to a normal distribution. Given a set of

data 1 2, , , Nx x x , the formula of kurtosis is:

( )

4

1

4ek 3

N

iix x N

s

=−

= −

(2)

where x is the mean, s is the standard deviation, and N is the

sample size. Positive excess kurtosis indicates a ‘heavy-tailed’

distribution while negative indicates a ‘light-tailed’ distribution.

3.3 Social-demographic Groups Analysis

This step intends to identify the passengers’ social-demographic

groups by clustering LTDS data. Comparing to classical

clustering tasks, LTDS data contain both continuous (e.g. age

and income) and categorical variables (e.g. main occupation).

TwoStep Cluster (Bacher et al. 2004) is a suitable algorithm to

deal with this clustering task. In addition, TwoStep algorithm is

a scalable cluster method, allowing to analyse large dataset and

it can automatically determine the optimal number of clusters.

TwoStep Cluster algorithm involves three main steps: pre-

clustering, outlier handling (optional) and clustering. The pre-

cluster step is implemented by building a modified cluster

feature tree. The clustering procedure is to group the sub-

clusters resulting from the pre-cluster step into an optimal or a

No. feature Description

Temporal

feature

AFTI_WD The average start time of the first trip on weekdays

LFTI_WD The average start time of the last trip on weekdays

AFTI_WE The average start time of the first trip on weekends

LFTI_WE The average start time of the last trip on weekends

MPT_TUBE_NUM the number of trips by tube during morning peak

EPT_TUBE_NUM the number of trips by tube during evening peak

MPT_BUS_NUM the number of trips by bus during morning peak

EPT_BUS_NUM the number of trips by bus during evening peak

MPTR_TUBE Morning peak travel rate by tube

EPTR_TUBE Evening peak travel rate by tube

MPTR_BUS Morning peak travel rate by bus

EPTR_BUS Evening peak travel rate by bus

SEASON_1/2/3/4 The number of trips during the 1/2/3/4-th season

SEA_PER_1/2/3/4 The percentage of trips during the 1/2/3/4-th season

Spatial

Features

AVG_T_WD The average of tube trip time on weekdays

AVG_T_WE The average of tube trip time on weekends

VAR_T_WD The variance of tube trip time on weekdays

VAR_T_WE The variance of tube trip time on weekends

AVG_MAX_TD The average radius travelled by tube per day

VAR_MAX_TD The average radius travelled by tube per day

TOTAL_TD The total travel distance by tube in the whole year

AVG_TS The daily average of the number of visited tube stations

VAR_TS The daily variance of the number of visited tube stations

AVG_BL The daily average of the number of visited bus lines

VAR_BL The daily variance of the number of visited bus lines

ZONE_T_R How often a passenger transfers the travel zone per day

AVG_INNER The mean value of the inner zone number

AVG_OUTER The mean value of the outer zone number

VAR_ZONE_IO The variance differences of inner-zone and outer-zone

Travel Mode

Features

TUBE_NUM The total number of the tube journeys

BUS_NUM The total number of the bus journey

TUBE_PER The percentage of tube journeys

Travel

Frequency

Features

TRA_DAY How many days a passenger travels in the whole year

TRA _DUR Travel duration in the whole year

TRA_WD How many weekdays a passenger travels in the whole year

TRA_WE How many weekends a passenger travels in the whole year

TRA_R_WD Weekday travel rate (TRA_WD/ TRA _DUR)

TRA_R_WE Weekend travel rate (TRA_WE/ TRA _DUR)

WD_TRIP The total number of weekday trips

WE_TRIP The total number of weekend trips

AVG_WD_TRIP The average number of weekday trips per day

AVG_WE_TRIP The average number of weekend trips per day

Table 1. Feature extracted from smart card data



1377

desired number of clusters. This process is implemented by

using the hierarchical clustering algorithm, which can produce a

sequence of partitions in one run. To determine the optimal

cluster solutions, each potential number of clusters is compared

using Schwarz's Bayesian Criterion (BIC) or the Akaike

Information Criterion (AIC) as the clustering criterion.

3.4 Association analysis

The discussion of the relationship between travel patterns and

social demographics are still somewhat ambiguous in existing

literature. Previous work usually summarised the demographic

attributes based on the results of travel pattern segmentations

(Ortega-Tong 2013), which can be regarded as a one-to-one

relationship mode. A more reasonable assumption of the

relationship between travel patterns and social-demographics

should be a many-to-many mode considering the following

reasons.

First, the previous passenger segmentation totally depends on

the individual or household social-demographics. The

segmentation results cannot reflect whether the selected social-

demographic characteristics are indeed significant determinants

of travel patterns at the individual level. Secondly, according to

previous researches, some social-demographic characteristics,

such as age, income, and car ownership, can largely affect

personal travel patterns. However, the complex travel

behaviours are not determined by a single demographic feature,

but the combination of diverse social-demographic attributes, as

well as some other unknown latent factors.

To achieve a better explanation of the individuals’ complex

travel patterns, we need to find more significant relationships

between travel patterns and the social-demographic

characteristics while keeping the diversity of travel patterns to

the largest extent. Thus, we aggregate the initial social-

demographic categories by applying hierarchical clustering

(HC) (Kraskov et al. 2005).

The third step is based on the results of the first two steps. After

we obtained the travel patterns in the first step and the

demographic groups in the second step, it is found that people

in the same demographic group may exhibit different travel

patterns. Thus, we use the distribution of passengers over

different travel patterns as the feature vector of each

demographic group, as illustrated in Figure 2. For example, the

demographic group 1 has 50% of passengers exhibit the second

travel pattern and 11% of passengers exhibit the M-th travel

pattern, as shown in Figure 2. Using these feature vectors, HC

clustering is then applied to aggregate demographic groups to

identify significant relationships between demographic groups

and travel patterns. HC starts by treating each observation as a

separate cluster. Then, it repeatedly executes the following two

steps: (1) identify the two clusters that are the closest together,

and (2) merge the two most similar clusters. This continues until

all the number of clusters are equal to the predetermined value.

This is illustrated in the diagrams below. To determine the

optimal number of clusters, we use the Dunn Index to measure

the clustering performance.

4. CASE STUDY

In this section, we use London’s Oyster Card and LTDS data

from 9708 passengers to demonstrate the proposed framework

of exploring the relationship between travel patterns and social-

demographics. Details are given bellow.

Figure 2. The feature vector of each demographic group is the

passenger distribution across different travel patterns

4.1 Travel Patterns of Residents in London

4.1.1 Data pre-processing

Travel features of 9708 passengers are extracted as described in

section 3.2.1. Before clustering, features should be first rescaled

to remove the influence of the different data range. Second, the

extracted features include spatial, temporal, mode preference

and travel frequency characteristics. Since the dimension of the

travel measures is large and some of them are intercorrelated,

PCA is applied to reduce the dimensionality. The number of

principal components to be retained is automatically estimated

by using the method proposed by Minka (2000). Finally, the

first 20 components are kept, explaining around 96.8% of the

total variance.

4.1.2 Travel Pattern Clustering Results

AP is used to detect travel pattern clusters. We calculate the

Dunn index by running AP with the different number of

predefined clusters ranging from 2 to 20. According to Figure 3,

the Dunn index reaches the local maximum value at 15 clusters,

indicating the optimal segmentation.

Figure 3. The Dunn index changes with the number of clusters

obtained by Affinity Propagation algorithm

The 9708 passengers are classified into 15 clusters, as shown in

Figure 4. The largest group contains around 14% passengers

while the smallest cluster (cluster 15) only consists of 94

passengers (less than 1%). Observing the travel features of

cluster 15, we find over 95% of individuals in this cluster only

used their Oyster cards once or twice during the whole year.

Thus, we think these Oyster cards are just for disposable use

and we do not further discuss it.

Figure 4. The size of each travel pattern



1378

4.1.3 Seasonality identification

For the rest of 14 clusters, we would like to further identify the

seasonal travel patterns. The values of skewness of the 14

clusters are listed in Table 2. Referring to the above-mentioned

rules, we summarise the 14 clusters into two main categories,

unseasonal travel patterns (cluster 1 to 7) and seasonal travel

patterns (clusters 8 to 14). Figure 5 illustrates three distinct

distribution of trips over the overall year. Overall, the number

of unseasonal passengers in the first seven clusters is 5515,

accounting for around 53.1% of the total population, and

seasonal passengers are 4533 (near 46.9%), a little fewer than

the unseasonal. Since the second step for seasonality

quantitation is only applied to seasonal travel patterns, we move

this part to the semantic analysis in the next subsection 4.1.4.

Unseasonal travel patterns

Cluster No. 1 2 3 4 5 6 7

skewness -0.06 -0.03 -0.19 -0.20 -0.12 -0.29 -0.16

Seasonal travel patterns

Cluster No. 8 9 10 11 12 13 14

skewness -1.05 -1.17 -1.51 -1.08 -0.92 -0.61 0.53

Table 2. The values of skewness of the 14 clusters

Figure 5. The distribution of trips of Cluster 2, 10 and 14 during

the year 2012

4.1.4 Semantic analysis of travel patterns

(1) Unseasonal Passengers

Clusters 1 to 7 denote unseasonal daily routine travel

behaviours. According to Table 3, a first similarity can be made

among these clusters: clusters exhibiting relative long travel

duration. Then, the distinct travel frequencies and preferred

travel modes identify varying types of typical unseasonal travel

patterns.

• Cluster 1/2: Unseasonal heavy bus/tube users

There are 378 passengers in Cluster 1 identified as the most

frequent usage of bus services and the longest travel duration

(number of days before the first and the last day on which using

public transit). According to Table 3 their first and last daily

trip times are relatively early among all clusters. And these

passengers took public transit over 4 times per day. In addition,

Cluster 2 shows quite similar travel frequency, travel days and

durations. The difference between Cluster 1 and 2 is the

opposite travel mode preference. In addition, on average, users

in this cluster have the earliest mean time of the first journey

and the latest mean time of the last trips on weekdays. These

attributes strongly imply the purpose of commuting.

• Cluster 3/4/5: Unseasonal moderate bus/tube/mixed-

mode users

The second subgroup contains three clusters consisting of

unseasonal, moderate public transit users with different travel

mode preferences. Specifically, Cluster 3 represents passengers

who usually travel by bus. However, most trips occurred during

off-peak time. Passengers in Cluster 4 exhibit very similar

temporal behaviours with Cluster 2, but the evening use of

Cluster 4 is around one-hour earlier than that of Cluster 2. In

addition, passengers in Cluster 4 travel more on weekdays than

any other unseasonal type. Comparing to Cluster 3 and 4,

Cluster 5 has the latest first and last departure time. The travel

mode is somewhat irregular.

• Cluster 6/7: Unseasonal occasional bus/mixed-mode

users

Regarding the remaining two Clusters 6 and 7, the most

common features are the long travel duration but few and

diffuse travel days. Additionally, the proportions of trips during

rush hours of the two clusters are both lower than 15%, but

their travel modes are different. The former prefers to use bus

while the latter has no obvious preference. On weekends,

residents in Cluster 7 made more evening trips than Cluster 6.

Another significant distinction exists in the spatial features. The

range of motion of passengers in Cluster 6 aims at the inner

city, and Cluster 7 is the opposite.

(2) Seasonal Passengers

Clusters 8 up to 14 are identified as seasonal travel patterns,

thus we compute the excess kurtosis to identify their season

preference. Results are listed in Table 4. We can see that the

first 4 clusters are heavy-tailed, and the rest are light-tailed, and

we also point out the favourite travel season of each cluster (the

fifth row in Table 4).

In addition, passengers in the first seven clusters have similar

seasonal behaviours, trending of which is like the subplot (b) in

Figure 5. Only the last cluster No.14 shows the opposite trend

like Figure 5 (c). The similarity is that all the most-frequent

travel periods are in the winter, indicating the influence of

seasonality on residents’ travel behaviour. More details of the

semantic analysis of each travel patterns are given as follows.

• Cluster 8: Seasonal heavy bus users

These passengers heavily rely on bus for daily trip, but the trip

distribution is seasonal. Passengers averagely use the public

transit for about 92 days out of the 113-day travel duration,

which results in a very dense bus travel demand. With regards

to daily temporal behaviour, people in this cluster travel earlier

in the morning and later in the evening than other seasonal

patterns on weekdays. In spatial respect, according to these

passengers’ sparse tube journeys, the average tube travel zones

reveal that they only use tube very far away from central

London.

• Cluster 9/10/11: Seasonal moderate bus/tube/mixed-

mode users

Passengers in Clusters 9 to 11 show a moderate travel frequency

with distinct travel modes (bus, tube and mixed, respectively)

during winter. Among them, Cluster 9 and 11 exhibit a similar

temporal behaviour during weekdays. The significant features

of Cluster 9 are the extremely short average travel distance and

narrow travel zone by tube. In addition, Cluster 10 exhibits a

remarkable temporal similarity with Cluster 8 on weekdays.

Comparing with other seasonal travel patterns, another

considerable feature of Cluster 10 is the high proportion of

weekday trips, which proportioned for over 75% of the total

number of trips. In addition, approximate a quarter of trips

occurred during morning and evening peak hours. These

features strongly indicate Cluster 10 is a typical seasonal travel

pattern with a main purpose of commuting.

• Cluster 12/13/14: Seasonal occasional bus/tube/mixed-

mode users



1379

These passengers rarely travel by public transit, and their

infrequent travels always occurred during a short period.

Clusters 12 and 14 have quite similar temporal behaviours

(except seasonality) that they both exhibit a late morning usage

at about 12:00 and early evening ending before 16:00 on

weekdays, and they hardly use public transit during weekends,

therefore the trips on weekdays take account for almost 100% of

the total trips. In terms of Cluster 13, it shows the longest travel

duration among all seasonal travel patterns, resulting in a more

diffuse usage.

4.2 Social-demographic groups

In this case study, ten socio-demographic features collected in

LTDS are considered for clustering. In LTDS, age (PAGEI) and

the distance between home and work/education (PLENN) can

be treated as continuous variables and others are categorical

variables.

In the TwoStep clustering, the BIC is calculated as the

clustering performance metric to determine the optimal cluster

number. As smaller values of BIC indicating better models, 32

clusters are chosen as the most efficient and practical number,

preserving a significant diversification of the residents in

London. For privacy reason, we cannot provide the details of

the social-demographic characteristics for each group. We only

present the 32 clusters’ average social-demographic features

ordered by the average age in Figure 6. To achieve a better

visualisation, each demographic feature has been scaled by the

maximum value.

Figure 6. Sankey plot of 32 social-demographic groups

In summary, the first three demographic groups can be regarded

as teenagers under education. The dominant distinctions are

Unseasonal passengers

Heavy Moderate Occasional

Cluster No. 1 2 3 4 5 6 7

First/Last tap-in time on weekdays 10:20

17:30

9:10

18:40

11:30

15:50

9:10

17:20

11:50

17:20

12:10

14:15

12:30

15:25

First /Last tap-in time on weekends 12:00

16:40

12:50

17:40

12:40

16:00

12:50

16:30

13:40

17:40

12:00

13:50

13:30

16:10

Tube trip rate in morning/evening peak (%) 2.0/2.6 24.9/22.1 2.7/3.5 29.0/26.8 3.1/8.2 0.5/0.8 4.2/6.3

Bus trip rate in morning/evening peak (%) 14.0/15.1 6.5/1.2 16.2/13.2 6.4/5.2 11.2/11.6 13.1/10.2 9.6/8.9

Tube travel zones 1.68-2.98 1.41-2.77 1.73-3.18 1.41-3.32 1.69-3.27 1.36-1.39 1.69-3.69

Mean distance of tube trip (km) 4.95 5.04 4.97 5.96 5.51 0.42 6.54

Mode preference Bus Tube Bus Tube Mix Bus Mix

Total tube/bus trips 115/998 567/169 62/361 303/102 98/126 2/87 43/68

Weekdays percentage 59.42% 63.53% 63.96% 82.34% 64.12% 63.96% 62.64%

Travel days 266 256 147 149 98 38 36

Travel duration 344 339 307 292 198 206 273

Table 3. Selected travel features of unseasonal travel patterns

Seasonal passengers

Heavy Moderate Occasional

Cluster No. 8 9 10 11 12 13 14

Excess Kurtosis value 0.458503 0.626613 1.382376 0.444484 -0.23215 -0.96314 -0.59545

Favourite season 4 4 4 4 4 4 1

First/Last tap-in time on weekdays 10:40

17:40

11:10

15:40

10:40

17:40

12:25

16:10

12.25

14:10

13:15

16:50

12:10

15:30

First /Last tap-in time on weekends 12:10

17:30

12:20

15:30

10:40

13:50

13:30

16:40

-

-

13:50

17:15

-

-

Tube trip rate in morning/evening peak (%) 4.0/4.7 0.5/0.7 22.4/22.6 7.7/4.6 0.4/0.8 12.2/22.8 11.4/14.8

Bus trip rate in morning/evening peak (%) 13.9/12.8 18.1/12.5 6.7/5.5 10.1/8.5 14.0/11.8 1.9/1.9 6.8/10.0

Tube travel zones 1.79-3.26 1.45-1.49 1.42-2.93 1.74-3.39 1.15-1.17 1.49-3.28 1.48-3.66

Mean distance of tube trip (km) 5.13 0.53 5.25 5.51 0.22 5.70 6.60

Mode preference Bus Bus Tube Mix Bus Tube Mix

Total tube/bus trips 74/314 2/172 83/28 43/61 0.6/16 51/12 16/12

Weekdays percentage 55.29% 58.52% 75.45% 59.34% 99.76% 51.46% 99.76%

Travel days 92 52 43 38 8 28 12

Travel duration 113 78 66 104 44 175 135

Table 4. Some selected travel features of seasonal passengers



1380

their household characteristics, including income, the distance

between home education place, and car ownership, as well as

two personal characteristics (car driver/passenger frequency).

Then, we treat groups 4 up to 23 as middle-aged adults, which

exhibit the most diverse demographic features at both

household and individual level. Among them, group 6, 9, and

14 are unemployed. Finally, the rest 9 groups (from 24 up to

32) mainly consist of retired old-age people grouped by using

the household characteristics.

4.3 Significant relationship analysis

Passengers’ travel behaviours strongly depend on their

demographics. However, because of some unknown factors,

such as subjective travel preference, and the accessibility of PT,

individuals in the same demographic groups may exhibit

different travel patterns. However, comparing the passenger

distribution across the travel patterns, we find that some

demographic groups presented a quite similar distribution.

Thus, Hierarchical Clustering is applied to this distribution to

aggregate original social-demographic groups. The aggregation

process and the relationship between the aggregated

demographic groups and the travel patterns are presented using

a flowchart in Figure 7.

Figure 7. The aggregation process of 32 demographic groups

and the relationship between demographic groups and travel

patterns. The left denotes the original 32 social-demographic

groups, the middle is the 9 groups aggregated by the passenger

distribution across travel patterns, and the right are the 15 travel

patterns.

To further explain the semantic meaning of the aggregation

results, we selected two typical examples to give more details of

the semantic analysis of the relationship.

(1) Young passengers

Young passengers in the first three original social-demographic

groups are merged together as group 1 in Figure 7 in this

aggregation process. Observing the passenger distribution

across travel patterns, almost half of them (total 807 persons)

belong to travel pattern 5, which is described as unseasonal

moderate mixed-mode travellers. It means that young

passengers, most of whom are students, have no obvious

preference for a certain travel mode. What’s more, because the

working time is not as fixed as office workers, they did not

always travel during the morning peak.

(2) Old passengers

The old passengers are merged into two groups, group 4 and

group 6 in Figure 7, respectively. The former is combined of

demographic groups 30 to 32 (the oldest three) and the latter

includes demographic groups 24 to 29. The average social-

demographic characteristics of the two aggregated groups are

presented using a radar plot in Figure 8. It can be seen in Figure

8, passengers in group 4 (average 74-year-old, 879 people) are

slightly older than those in group 6 (average 61-year-old, 2037

people). In addition, although the working status of the two

groups indicate that most of the people are retired the average

levels of car ownership, household income and frequency of car

driver of group 4 are considerably lower than that of group 6.

The passenger distribution of the two groups across 15 distinct

travel patterns can be seen in Figure 9. For group 4, a

significant feature is that most of the oldest prefer to use bus

(travel pattern 1, 3, 6, 8, 9 and 12) and over 60 % passengers in

group 4 exhibit unseasonal patterns, which implies that their

daily mobility highly depends on the public transport, especially

bus system. The potential reasons include the cheap ticket

prices and no demand for commuting. The travel mode

preference of group 6 is similar to group 4. However, the tube

usage of group 6 is more frequent than group 4.

group 4 group 6

Figure 8. The average demographics of group 4 and 6.

Figure 9. The passenger distribution of group 4 and 6 across 15

distinct travel patterns

5. CONCLUSIONS AND FUTURE WORK

Smart card data provide a promising opportunity to investigate

the complex travel behaviours in public transport system. This

paper proposes a novel and entire framework to analyse the

significant relationships between travel patterns and social-

demographics of passengers using smart card data and

household survey. This effort provides some new insights into

the spatio-temporal travel patterns and their linkage between

demographic roles of passengers.

Future work can be conducted based on the research presented

in this paper. First, the extracted features from SC data can

reveal travel behaviours from the spatial, temporal, travel mode

and frequency perspectives, but each feature is just the mean

value during the research period, which may miss some useful



1381

behaviour features for travel patterns analysis. Thus, other more

effective methods should be explored to represent the SC data

to describe the travel behaviour of passengers. Second,

exploring the possibility of predicting social-demographic roles

using SC data is an interesting feature research direction.

ACKNOWLEDGEMENTS

Transport for London is acknowledged for sharing London’s

Oyster card data and LTDS data, which are used in this study.

This work was supported in part by the Consumer Data

Research Centre, Economic and Social Research Council, U.K.,

under Grant ES/L011840/1. The first author’s research is jointly

supported by the China Scholarship Council under Grant No.

201603170309 and the Dean's Prize from the University

College London.

REFERENCES

Bacher, J., Wenzig, K. and Vogler, M., 2004. SPSS TwoStep

Cluster-a first evaluation.

Bulmer, M., 1979. Principles of Statistics Dover Publications.

New York.

Ceapa, I., Smith, C. and Capra, L., 2012. Avoiding the crowds:

understanding tube station congestion patterns from trip data.

Proceedings of the ACM SIGKDD international workshop on

urban computing. ACM.

Frey, B. J. and Dueck, D., 2007. Clustering by passing

messages between data points. science 315 (5814), 972-976.

Kieu, L. M., Bhaskar, A. and Chung, E., 2014. Transit

passenger segmentation using travel regularity mined from

Smart Card transactions data.

Kraskov, A., Stögbauer, H., Andrzejak, R. G. and Grassberger,

P., 2005. Hierarchical clustering using mutual information. EPL

(Europhysics Letters) 70 (2), 278.

Lathia, N. and Capra, L., 2011. How smart is your smartcard?:

measuring travel behaviours, perceptions, and incentives.

Proceedings of the 13th international conference on Ubiquitous

computing. ACM.

Minka, T. P., 2000. Automatic choice of dimensionality for

PCA. Nips. Vol. 13.

Nassir, N., Hickman, M. and Ma, Z.-L., 2015. Activity

detection and transfer identification for public transit fare card

data. Transportation 42 (4), 683-705.

Ortega-Tong, M. A., 2013. Classification of London's public

transport users using smart card data. Massachusetts Institute

of Technology.

Pelletier, M.-P., Trépanier, M. and Morency, C., 2011. Smart

card data use in public transit: A literature review.

Transportation Research Part C: Emerging Technologies 19

(4), 557-568.

Sari Aslam, N., Cheng, T. and Cheshire, J., 2018. A high-

precision heuristic model to detect home and work locations

from smart card data. Geo-spatial Information Science, 1-11.

Shi, X. and Hangfei, L. (2014) The analysis of bus commuters’

travel characteristics using smart card data: the case of

Shenzhen, China.

Sun, L., Lee, D.-H., Erath, A. and Huang, X., 2012. Using

smart card data to extract passenger's spatio-temporal density

and train's trajectory of MRT system. In Proceedings of the

ACM SIGKDD International Workshop on Urban Computing.

Beijing, China. 2346519: ACM. 142-148.

Tao, S., Rohde, D. and Corcoran, J., 2014. Examining the

spatial–temporal dynamics of bus passenger travel behaviour

using smart card data and the flow-comap. Journal of Transport

Geography 41, 21-36.

Yuan, N. J., Wang, Y., Zhang, F., Xie, X. and Sun, G., 2013.

Reconstructing Individual Mobility from Smart Card

Transactions: A Space Alignment Approach. 2013 IEEE 13th

International Conference on Data Mining. 7-10 Dec. 2013.

Zhang, Y. and Cheng, T., 2017. Feature Extraction for Long-

term Travel Pattern Analysis. Proceedings of the 25th GISRUK

conference. Manchester, UK.

Zhang, Y. and Cheng, T., 2018. Inferring Social-Demographics

of Travellers based on Smart Card Data. 2nd International

Conference on Advanced Research Methods and Analytics.

Valencia, Spain. Editorial Universitat Politècnica de València.

Zhang, Y. and Cheng, T., 2019. A Deep Learning Approach to

Infer Employment Status of Passengers by Using Smart Card

Data. IEEE Transactions on Intelligent Transportation Systems.



1382

Documents

EXPLORING THE RELATIONSHIP BETWEEN TRAVEL PATTERN …€¦ · EXPLORING THE RELATIONSHIP BETWEEN TRAVEL PATTERN AND SOCIAL-DEMOGRAPHICS USING SMART CARD DATA AND HOUSEHOLD SURVEY